Vowel analysis using FastTrack

Fast Track is a (nearly) automated formant tracking tool that enables you to extract the formant contours quickly and easily. The tool functions based on a number of praat scripts and as a plug-in in Praat, so there is no need for you to code anything. There is also room for customising the scripts to make more individualised tool. Fast Track is developed by Santiago Barreda at UC Davies and more information can be found here: https://github.com/santiagobarreda/FastTrack

Also, the following paper explains the tool in more detail:

Barreda, S. (2021). Fast Track: fast (nearly) automatic formant-tracking using Praat. Linguistics Vanguard, 7(1), 20200051. https://doi.org/10.1515/lingvan-2020-0051

1 Preliminaries

In the workshop, I will explain and demonstrate the following things:

1. The fun bit: Bulk formant estimation
2. Another fun bit: Tidying up data for R
3. Slightly tedious bit: Extracting vowels

If you would like to follow along, you can install FastTrack beforehand. A detailed step-by-step guide is available in Santiago’s Github repository with some video illustrations. See the wiki on his Github repository for the tutorial on installation (and many other things!)

Prepared data

I have prepared data that are compressed into zip files. Each corresponds to different stages I plan to go through during the demonstration:

• NWS.zip:
  • This is basically a folder that contains WAV and TextGrid files, but are stored in separate folders in a FastTrack-friendly manner. This is the folder structure required to extract vowels.
• NWS_individual_vowels.zip:
  • The folder contains the product of the vowel extraction. Therefore, each sound file corresponds to a vowel and is stored in the ‘sounds’ folder. There are also two csv files that are helpful for tidying up the data afterwards.

Speech data

For demonstration, I’m going to use recordings of the well-known passage “The North Wind and the Sun” that are publicly available in the ALLSTAR Corpus. The ALLSTAR Corpus contains a number of spontaneous and scripted speech that were produced by English speakers from different language backgrounds.

Bradlow, A. R. (n.d.) ALLSSTAR: Archive of L1 and L2 Scripted and Spontaneous Transcripts And Recordings. Retrieved from https://oscaar3.ling.northwestern.edu/ALLSSTARcentral/#!/recordings.

2 Bulk formant estimation

Let’s imagine that you are interested in comparing vowel realisations between a native English speaker and a native Japanese learner of English.

The highlight of using FastTrack is that it enables you to extract formant frequencies from multiple files automatically. So, skipping all the preparation, here are what you can expect from FastTrack:

• FastTrack does a lot of regressions and chooses the best analysis out of multiple candidates automatically. It can also return images of all candidates and the winners for visual inspection.

• It estimates the formant frequencies at the multiple time points throughout the vowel duration.

• The output is a csv file summarising the analysis, which can then be imported into R for tidy up, visualisation, statistics, etc.

You can extract formant contours easily through the following steps:

1. Download NWS_individual_vowels.zip, unfold it and save it somewhere on your computer.

2. Open Praat and throw a random file in the object window. This will trigger the FastTrack functions to appear in the menu section (if FastTrack is installed properly).

3. Select Track folder….

4. Specify the path to the NWS_individual_vowels folder in the “Folder” section. You can also adjust some other functions if prefer. For more details, please consult Sandiego’s Github page.

• Note that the path to be specified here is the folder that contains a ‘sound’ folder and two other csv files. This won’t work if you specify the ‘sound’ folder directly.

• Also note that the naming convention is also important in later stages - i.e. a ‘sounds’ folder and a ‘texrgrids’ folder.

Figure 2.1: Setting window for formant estimation (left) and an example of comparison image (right)

This operation yields quite a few output files, for which we don’t have enough time to look through. If interested, you can download “NWS_results.zip” to see the full output. The most important here is the results of the formant estimation, stored in a csv file as aggregate_data.csv in processed_data folder. Here is a glimpse of it:

Figure 2.2: An example of Aggregate_data.csv

3 Tidying up data for R

I think FastTrack becomes even more powerful if combined with R. This, however, requires you to figure out how to tidy up the data for a more R-friendly data frame. If interested, you can look at my R code (FastTrack.R) to tidy up the data, which I wrote based on Sam’s script.

Below is my attempt to visualise vowels of nonnative English speakers from the Asian countries (Hong Kong, Thailand and Vietnam) as well as an American English speaker.

Figure 3.1: Comparison of vowel space in Asian Englishes speakers and an American English speaker

4 Extracting vowels

OK, we have seen that FastTrack is a very powerful tool in automatic formant estimation for a bulk of vowels and then R could also be used to play around with the FastTrack data. However, formant estimation requires each audio file to contain vocalic components throughout the duration, which means that we need to clip out every vowel token from the recording. Thankfully, FastTrack is also ready to do this for us.

If you wish to follow along, the following instructions should work:

1. Download NWS.zip and save it somewhere on your computer. Again, I have already made the structure optimal for the vowel extraction.

2. Open Praat and throw a random file in the object window. This will trigger the FastTrack functions to appear in the menu section.

3. Select Tools…, then Extract vowels with TextGrids.

4. Specify the following:

• Sound folder:
  • Path to “sounds” in the NWS folder containing .wav files.
• TextGrid folder:
  • Path to “textgrids” in the NWS folder containing .TextGrid files.
• Output folder:
  • Path to the folder where you wish to save the outputs. You could specify an existing location.
• Which tier contains segmentation information?:
  • Specify the tier in which phonemic transcription/segmentation has been performed. In the current example, the segmentation is done in Tier 2 so type 2.
• Which tier contains word information?:
  • Specify the tier with words. Type 1 in this case.
• Is stress marked on vowels?:
  • Tick the box if you wish to take stress into account. If you use forced alignment to segment the speech, stress is marked alongside each vowel. For example, you will find “AE1”, which means a TRAP vowel that bears the primary stress, or “AE2” the secondary stress, etc.

Figure 4.1: Vowel extraction setting window

5 Other interesting stuff

I have demonstrated the overview of my vowel analysis workflow, but these are only basic things.

1. Segmenting audio files into phonemes using Montreal Forced Aligner

   • You need audio files and acoomanying Praat TextGrids that contain a tier with phonemically segmented intervals and I find it useful to use Montreal Forced Aligner (MFA) to prepare the files. Dr Eleanor Chodroff recently held a workshop on using MFA at Rutgers University and the video can be found on Youtube https://www.youtube.com/watch?v=Zhj-ccMDj_w. She also develops a comprehensive user manual for MFA which can be found at https://eleanorchodroff.com/tutorial/montreal-forced-aligner-v2.html.

2. Formant correction

   • At this stage, my analyses remain very coarse as it still takes me a lot of time checking if the segmenetation is correct before extracting vowels. However, you may want to check the accuracy of formant estimation and hand correct the measurement if necessary. I haven’t managed to learn it just yet, but FastTrack has a capacity for this. You can learn about it in How to analyze a folder and Fixing errors.